A July 14, 2026, Google Developers Blog post details a systems engineering playbook for optimizing the Qwen 3.5-397B MoE model on the Ironwood (TPU v7x) platform. The model has 397 billion total parameters, activating exactly 17 billion parameters per token per forward pass, representing a 4.3% routing activation ratio. The team applied a modular, model-agnostic optimization strategy using reusable JAX/Pallas kernels, targeting novel components like Gated DeltaNet (GDN) linear attention and Attention Data Parallelism. Between April and June 2026, these optimizations improved inference performance by approximately 3.1x for Decode-heavy and by approximately 4.7x for Prefill-heavy workloads (512 Concurrency tier). The optimizations were integrated into open-source serving frameworks like vLLM and SGLang to provide a production-ready migration path for enterprise workloads.